keep it simple? predicting primary health care costs with...

This is a repository copy of Keep it Simple? Predicting Primary Health Care Costs with Measures of Morbidity and Multimorbidity.

White Rose Research Online URL for this paper:http://eprints.whiterose.ac.uk/136659/

Version: Published Version

Monograph:Brilleman, S. L., Gravelle, Hugh Stanley Emrys orcid.org/0000-0002-7753-4233, Hollinghurst, S. et al. (3 more authors) (2011) Keep it Simple? Predicting Primary Health Care Costs with Measures of Morbidity and Multimorbidity. Discussion Paper. CHE Research Paper . Centre for Health Economics, University of York , York, UK.

[email protected]://eprints.whiterose.ac.uk/

Reuse Items deposited in White Rose Research Online are protected by copyright, with all rights reserved unless indicated otherwise. They may be downloaded and/or printed for private study, or other acts as permitted by national copyright laws. The publisher or other rights holders may allow further reproduction and re-use of the full text version. This is indicated by the licence information on the White Rose Research Online record for the item.

Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing [email protected] including the URL of the record and the reason for the withdrawal request.

mailto:[email protected]

https://eprints.whiterose.ac.uk/

Keep it Simple? Predicting Primary Health

Care Costs with Measures of Morbidity and

Multimorbidity

CHE Research Paper 72

Keep it Simple? Predicting Primary Health Care

Costs with Measures of Morbidity and

Multimorbidity

1Samuel L Brilleman2Hugh Gravelle1Sandra Hollinghurst1Sarah Purdy1Chris Salisbury3Frank Windmeijer

1Academic Unit of Primary Care, University of Bristol2Centre for Health Economics, University of York3Department of Economics, University of Bristol

December 2011

Background to series

CHE Discussion Papers (DPs) began publication in 1983 as a means of making currentresearch material more widely available to health economists and other potential users. So asto speed up the dissemination process, papers were originally published by CHE anddistributed by post to a worldwide readership.

The CHE Research Paper series takes over that function and provides access to currentresearch output via web-based publication, although hard copy will continue to be available(but subject to charge).

Acknowledgements

The paper presents independent research commissioned by the National Institute for HealthResearch School for Primary Care Research. The views expressed are those of the authorsand not necessarily those of the NHS, the NIHR or the Department of Health. Helpfulcomments were received from Matt Sutton, and at the Society for Academic Primary Care2011 Bristol conference, and the Health Economists’ Study Group meeting in Bangor 2011.We are grateful to Tarita Murray-Thomas at GPRD for assistance and advice with the data.The project was covered by ethics approval for GPRD work granted by Trent MulticentreResearch Ethics Committee, reference 05/MRE04/87.

Disclaimer

Papers published in the CHE Research Paper (RP) series are intended as a contribution tocurrent research. Work and ideas reported in RPs may not always represent the final positionand as such may sometimes need to be treated as work in progress. The material and viewsexpressed in RPs are solely those of the authors and should not be interpreted asrepresenting the collective views of CHE research staff or their research funders.

Further copies

Copies of this paper are freely available to download from the CHE websitewww.york.ac.uk/che/publications/ Access to downloaded material is provided on theunderstanding that it is intended for personal use. Copies of downloaded papers may bedistributed to third-parties subject to the proviso that the CHE publication source is properlyacknowledged and that such distribution is not subject to any payment.

Printed copies are available on request at a charge of £5.00 per copy. Please contact theCHE Publications Office, email [email protected], telephone 01904 321458 for furtherdetails.

Centre for Health EconomicsAlcuin CollegeUniversity of YorkYork, UKwww.york.ac.uk/che

©Samuel L Brilleman, Hugh Gravelle, Sandra Hollinghurst, Sarah Purdy, Chris Salisbury,Frank Windmeijer

Keep it simple? Predicting primary heath care costs with measures of morbidity and multimorbidity i

Abstract

In this paper we investigate the relationship between patients’ primary care costs (consultations, tests,drugs) and their age, gender, deprivation and alternative measures of their morbidity andmultimorbidity. Such information is required in order to set capitation fees or budgets for generalpractices to cover their expenditure on providing primary care services. It is also useful to examinewhether practices’ expenditure decisions vary equitably with patient characteristics.

Electronic practice record keeping systems mean that there is very rich information on patientdiagnoses. But the diagnostic information (with over 9000 possible diagnoses) is too detailed to bepracticable for setting capitation fees or practice budgets. Some method of summarizing suchinformation into more manageable measures of morbidity is required. We therefore compared theability of eight measures of patient morbidity and multimorbidity to predict future primary care costsusing data on 86,100 individuals in 174 English practices. The measures were derived from fourmorbidity descriptive systems (17 chronic diseases in the Quality and Outcomes Framework (QOF),17 chronic diseases in the Charlson scheme, 114 Expanded Diagnosis Clusters (EDCs), and 68Adjusted Clinical Groups (ACGs)).

We found that, in general, for a given disease description system, counts of diseases and sets ofdisease dummy variables had similar explanatory power and that measures with more categories didbetter than those with fewer. The EDC measures performed best, followed by the QOF and ACGmeasures. The Charlson measures had the worst performance but still improved markedly on modelscontaining only age, gender, deprivation and practice effects.

Allowing for individual patient morbidity greatly reduced the association of age and cost. There was apro-deprived bias in expenditure: after allowing for morbidity, patients in areas in the highestdeprivation decile had costs which were 22% higher than those in the lowest deprivation decile.

The predictive ability of the best performing morbidity and multimorbidity measures was very good forthis type of individual level cross section data, with R

2ranging from 0.31 to 0.46. The statistical

method of estimating the relationship between patient characteristics and costs was less importantthan the type of morbidity measure. Rankings of the morbidity and multimorbidity measures werebroadly similar for generalised linear models with log link and Poisson errors and for OLS estimation.

It would be currently feasible to combine the results from our study with the data on the number ofpatients with each QOF disease, which is available on all practices in England, to calculate budgetsfor general practices to cover their primary care costs

Keywords: multimorbidity; primary care; utilisation; costs; deprivation; budgets

JEL categories: I14, I18

ii CHE Research Paper 72

Keep it simple? Predicting primary heath care costs with measures of morbidity and multimorbidity 1

1. Introduction

In this paper we investigate the relationship between patients’ primary care costs (consultations, tests,drugs) and their age, gender, deprivation and alternative measures of their morbidity andmultimorbidity. The relationship is of policy interest for two reasons. First, in many health caresystems general practitioners (GPs) are paid prospectively via capitation fees for the patients on theirlists. If policy makers are concerned either with equity, in the sense of ensuring that patients withgreater needs for health care carry a larger capitation, or with reducing financial incentives forpractices to cream skim or dump patients, then it is necessary to know how patients’ expectedexpenditure varies with their characteristics (Schokkaert et al., 1998). Second, such information canalso be used to investigate the equity of GPs’ decisions about the allocation of primary healthcareresources amongst their patients. Horizontal equity requires that patients in equal need shouldreceive equal amounts of health care. Vertical equity requires that there is greater expenditure onpatients in greater need. Models of expenditure can be used to test whether general practicesallocate primary care resources in accordance with vertical and horizontal equity: expenditure shouldbe related to morbidity and should not vary with a patient’s socio-economic status or the practice towhich they belong (Gravelle et al., 2006).

Whether one wishes to inform the setting of prospective capitation fees, or to test for equity, theimportance of morbidity as a determinant of costs implies that good measures of patient morbidity areessential for the empirical modelling of health care expenditure. With the development of electronicpatient record keeping systems in general practice it has become possible to obtain information ondetailed diagnoses for patients which can be used to construct such measures. The simplestapproach with such data is to group diagnoses into a manageable number of morbidity categorieswhich can then be included in regression models of patient costs as a set of dummy variablesindicating the presence or absence of the different diagnoses.

But measuring morbidity as a set of morbidity dummies does not allow for the possibility that the effectof diagnoses is not additive. The cost of one multimorbid patient with both diabetes and depressionmay be more than the cost of two patients, one with diabetes and the other with depression, becauseit may be more difficult to control blood sugar levels for a patient who also has depression.Conversely, there could be cost savings with some multimorbid patients. For example, heart diseaseand diabetes are conditions where monitoring of cholesterol may be required but the associated costsneed only be incurred once in a given period for a patient with both conditions. Allowing for thepossible non-additive effects of multimorbidity is potentially important since the proportion of thepopulation who are multimorbid is non-trivial (20% to 61% in our data set depending on themultimorbidity measure used) and has been growing over time (Hippisley-Cox and Pringle, 2007).

Multimorbidity measures range from simple counts of the number of morbidities to elaborateclassification schemes such as the widely used John Hopkins Adjusted Clinical Groups (ACG) Case-Mix System (Johns Hopkins Bloomberg School of Public Health, 2008). More elaborate schemesshould be better at predicting costs since they generally use more information and their constructionhas been guided in part by their predictive ability. But, particularly when used for setting capitationrates, simplicity is also a virtue. Simpler morbidity and multimorbidity schemes are easier for patientsand GPs to understand. Setting capitation fees based on morbidity requires that patient morbidity bemeasured every budgetary period for every patient and more complex schemes have highermeasurement and computation costs. Thus there may be a trade-off between simplicity andpredictive power when alternative measures of morbidity and multimorbidity are considered.

The main aim of this paper is to compare the ability of a set of morbidity and multimorbidity measuresof varying complexity to predict primary care costs. We believe it is the first to do so.

1

The paper makes a number of other contributions. First, our research is relevant for recent andproposed policy changes in the English NHS which will encourage patients to shop around amongstgeneral practices, with the lifting of restrictions on choice of practice (Department of Health, 2010) andwebsites such as NHS Choices providing information about practices. With greater mobility of patients

1Previous comparisons of the predictive power of alternative morbidity and multimorbidity schemes have focussed on hospital

cost (Huntley et al., 2012; Winkelman and Mehmud, 2008; Perkins et al., 2004). Omar et al (2008) examined prescribing costsfor English patients but did not compare morbidity measures.

2 CHE Research Paper 72

across practices it will be more important to align capitation funding of practices with the expectedcosts patients impose on practices. Capitation fees attached to patients are currently set by a formulabased on results from two models of the determinants of consultation rates and consultation length,as measures of GP workload. The study of consultation length did not include any morbidity variablesand the consultation rate study was based on area level data and used area measures of mortalityand of patient reported morbidity (Formula Review Group, 2007). Funding for general practiceprescribing is allocated by a formula derived from a practice level model of prescribing costs whichincluded three practice level disease prevalence measures (Department of Health, 2011). Weinvestigate whether it is possible to get accurate estimates of cost to be used in setting capitationpayments by using the data on individual patient morbidity which is readily available in practicerecords.

Second, there is some debate about estimation methods for health cost data, particularly about theusefulness of linear ordinary least squares models which do not take account of the skeweddistribution of costs and the high proportion of patients with zero costs (Manning, 2006). Thisliterature has focussed on modelling hospital costs whereas we estimate models of primary care cost.We examine whether the performance of alternative morbidity and multimorbidity measures issensitive to the estimation method.

Third, studies of equity in primary care have been based on population surveys which use a limitedrange of self reported measures of health (Bago, D’Uva, 2005; Bago D’Uva et al., 2009; Morris et al.,2005; Sutton, 2002). As a by-product of our main analysis we are able to test for horizontal equity byexamining whether expenditure varies with socio-economic status after controlling for a very muchricher set of morbidity measures which are based on data recorded by health care professionals.

Section 2 describes the data, the construction of the morbidity and multimorbidity measures, and theestimation methods. The results are set out in Section 3 and discussed in Section 4.


2. Methods

2.1 Data

2.1.1 Sample

The General Practice Research Database (GPRD) contains primary care medical records for around5 million patients currently registered throughout the United Kingdom (UK). An initial random sampleof patients aged 18 years and over was drawn from the 182 practices included in the GPRD whichhad ‘research standard’ data continuously from 1

stApril 2005 to 31 March 2008, and which had given

consent to link patient data to measures of area deprivation. The sample was stratified by age,gender and practice. The GPRD is considered broadly representative of the general population in theUK (Lawrenson et al., 1999). We dropped 8 practices with entirely missing deprivation data. In orderto use the most up-to-date resource use data and the largest possible observation period fordiagnoses, we included the 86,100 individuals from the original sample who were alive and registeredat one of the remaining 174 practices on 1

stApril 2007.

2.1.2 Costing

We calculated the total cost of primary care resources used by each patient during the NHS financialyear 1

stApril 2007 to 31

stMarch 2008. This included consultations, prescription drugs, and tests

initiated within primary care. All costs were valued in £ sterling at 2007/08 prices.

Consultations included all face-to-face (including surgery consultation, home visit, clinic, out of hours)and telephone consultations. The unit cost of each consultation was based on a combination ofconsultation type and primary staff role (type of general practitioner, practice nurse or other healthcare professional leading the consultation). The unit costs for consultations are shown in Table 1. Anadditional cost was attached to cover administrative activities such as the recording of results orsending mail to a patient when this was performed by a receptionist, administrator, or secretary. Unitcosts were taken from Curtis (2008) and from a report on GP earnings and expenses based on GPtax returns (Technical Steering Committee, 2010).

Table 1. Unit costs (£) per primary care encounter

Staff Type Surgeryconsultation

Home visit Clinic Telephoneconsultation

Out ofhours

GP: partner 24.47 81.37 35.98 14.85 36.97

GP: registrar/associate 15.92 52.92 23.40 9.66 24.05

GP: sole practitioner 27.55 91.61 40.51 16.72 41.63

Practice Nurse 9.00 - 9.00 5.46 -

Counsellor 64.00 - - - -

Other Health Care Professional 15.00 - 15.00 9.10 -

Unit costs for prescription drugs were based on information provided by the GPRD which combineddata from several sources, including the National Drug Tariff for generic products, and manufacturersfor branded products. Each prescription drug in the patient level data was matched to unit cost usingdrug name, strength and formulation. Where there was more than one unit cost for a prescriptiondrug we used the median unit cost.

To allow for possible ambiguities in the mapping of recorded drug quantities to costs we alsocomputed a measure of prescription costs using data from the prescription cost analysis of the NHSInformation Centre (2008a) which provides information on the net ingredient cost of all prescriptionsdispensed in the community in England. We used the British National Formulary (BNF) code (BritishMedical Association and Royal Pharmaceutical Society, 2010) to attach these average costs to eachprescription in the GPRD data, ignoring the quantity specified in the prescription.

With advice from a general practitioner member of the research team (SP) we determined which testswere performed within a standard surgery consultation and applied a zero unit cost, save for the costassociated with any consumables such as pregnancy test kits or urine dipsticks. Unit costs for theremaining tests were based on the National Health Service (NHS) Reference Costs (Department ofHealth, 2009). For laboratory tests the unit cost was based on pathology discipline. Hospital-basedtests and investigations requested by the practice were costed using the NHS Reference Costs.


2.1.3 Measures of morbidity and multimorbidity

We use four methods of categorising diagnoses to create three measures of morbidity and fivemultimorbidity measures (Table 2).

Table 2. Morbidity and multimorbidity measures

Measure Numberdiseases/categories

Range ofmeasure

Details

QOF disease dummyvariables

17Not mutuallyexclusive

0-1dummies

17 chronic diseases in the clinical domain of the UKQuality and Outcomes Framework (QOF) pay forperformance scheme. Not mutually exclusive

QOF disease count 17 0 to 17 Count of the QOF diseases.Charlson diseasedummy variables


0-1dummies

17 chronic diseases (similar to but not identical to theQOF set of chronic diseases). Not mutually exclusive.

Charlson Index score 17 0 to 33 A weighted score, where weights of 1, 2, 3, or 6 aregiven to each of the 17 chronic diseases in theCharlson set depending on the strength of theirrelationship with patient mortality.

Expanded DiagnosisClusters (EDCs)dummy variables


0-1dummies

EDCs are clinically related groupings of diagnosesselected by the ACG System. We include the 114 (ofthe possible 264) EDCs which have been identified aschronic. Not mutually exclusive.

Count of EDCs 114 0 to 114 Count of EDCsAdjusted ClinicalGroups(ACGs)

68 mutuallyexclusivecategories

0-1dummies

ACGs are mutually exclusive categories. Classificationinto an ACG is based on an individual’s age, gender,combination of morbidities and expected cost. Theage range of our sample meant we used only 68 outof 82 possible ACG categories.

Resource UtilizationBands (RUBs)

6 mutuallyexclusivecategories

0-1dummies

The ACG software groups ACGs into 6 mutuallyexclusive Resource Utilization Bands on the basis oftheir expected costs: 0 - No or only invalid diagnoses,1 - Healthy Users, 2 – Low, 3 – Moderate, 4 – High, 5,Very High. Patients assigned to higher bands areexpected to have higher costs.

QOF disease categories. The electronic record systems in UK general practices use Read codes torecord summary clinical information on patients. It is possible to use the Read codes to constructvarious sets of diagnostic categories.

2We used the 17 chronic conditions included in the clinical

domain of the 2006/7 version of the Quality and Outcomes Framework (QOF) which is a pay forperformance scheme covering all practices in the UK. This set of morbidity markers is simple andlikely to be reliably recorded because it forms the basis for payment under the QOF. It also has highface validity as the main business of general practices is dealing with chronic conditions, although itomits some chronic conditions such as skin disease and liver disease. The 17 conditions are asthma,atrial fibrillation, cancer, coronary heart disease (CHD), chronic kidney disease (CKD), chronicobstructive pulmonary disease (COPD), dementia, depression, diabetes, epilepsy, heart failure,hypertension, learning difficulties, mental health, obesity, stroke, and hyperthyroidism. We used theQOF Business Rules Version 16

3which outlines the clinical Version 2 Read codes and any additional

criteria required to include a patient on the relevant QOF disease register. The 17 QOF morbiditieswere included in the regression models as 17 dummy variables as description of patient morbidity.

In principle all possible types of multimorbidity can be represented by a suitable combination of thedummy variables for the separate morbidity categories. But with anything other than a very coarseset of categories, there will be too many combinations to be practical: with m different diagnosticcategories there are 2

mpossible morbidity states. Thus some means of aggregating the information

on diagnostic categories into more manageable descriptions of multimorbidity is required. Our firstmultimorbidity measure is a count of the number of QOF morbidity categories into which a patientfalls.

2It is possible to map from Read codes into ICD-10 which has over 9000 diagnostic categories in its finest classification and 20

in its coarsest (both counts excluding chapters XX and XXII). See http://www.who.int/classifications/icd/en/3

NHS Primary Care Contracting. QOF Implementation Business Rules v16. http://www.primarycarecontracting.nhs.uk/145.php


We drew on a review of the use of established indices of multimorbidity to consider which othermeasures might be appropriate to a UK primary care setting (Huntley et al, 2012). We selected theCharlson Index and measures derived from the John Hopkins ACG system because they are widelyused and potentially straightforward to operationalise with routine data.

Charlson diseases. The Charlson Index is a diagnosis-based multimorbidity measure that weights 17diseases on the basis of the strength of their association with mortality (Charlson et al., 1987). Weuse an adaptation by Khan et al (2010). The diseases included in the Charlson Index score, withassociated weight in parentheses, were cerebrovascular disease (1), chronic pulmonary disease (1),congestive heart disease (1), dementia (1), diabetes (1), mild liver disease (1), myocardial infarction(1), peptic ulcer disease (1), peripheral vascular disease (1), rheumatological disease (1), cancer (2),diabetes with complications (2), hemiplegia and paraplegia (2), renal disease (2), moderate or severeliver disease (3), AIDS (6), and metastatic tumour (6). Khan et al. (2010) provides the clinical Version2 Read codes for diagnosing each disease, using a translation from the widely used Deyo adaption ofthe Charlson Index for ICD-9 codes (Deyo et al., 1992). About half of the 17 conditions are similar tothose in the set of QOF chronic conditions, though the precise definitions vary. We used theweighted sum of the Charlson diseases (the Charlson Index) as a multimorbidity measure and avector of 17 dummy variables for the Charlson diseases to describe patient morbidity.

The John Hopkins ACG Case-Mix System is also diagnosis-based and was developed usingadministrative claims data in the United States (US) (Starfield et al. 1991; Weiner et al. 1991). Thesystem has been validated in a number of studies in the US and has been studied in primary care in,for example, Sweden (Halling et al. 2006) and Canada (Reid et al., 2002). Further studies have usedthe ACG system to predict hospital referrals and prescribing rates in the UK (Sullivan et al., 2005;Omar et al., 2008). The ACG system has recently been expanded to allow the input of Version 2Read codes. The software provides a variety of morbidity measures and multimorbidity measures.

Expanded Diagnosis Clusters (EDCs) are groupings of diagnostic codes which are clinically similar.An individual was assigned to an EDC if any of the Version 2 Read diagnostic codes relating to thatEDC appeared in their clinical data. We designated 114 of the 264 EDCs as representing a chroniccondition (Salisbury et al., 2011) and measured morbidity as a vector of 114 dummy variables. Wemeasured multimorbidity by the number of chronic EDCs an individual was included in. We used theJohn Hopkins ACG System Version 8.2 (Johns Hopkins Bloomberg School of Public Health, 2008) toobtain the EDC classifications.

Adjusted Clinical Groups (ACGs) are mutually exclusive categories (unlike the QOF, Charlson, andEDC categorisations). Diagnoses are first grouped into 32 non-mutually exclusive AggregatedDiagnosis Groups (ADGs) by duration, severity of condition, diagnostic certainty, etiology (infectious,injury, or other), and specialty care involvement. (EDCs are more purely diagnosis based.) ADGs arein turn grouped into 12 collapsed ADGs (CADGs) based on persistence, severity, type of healthcarerequired. ACGs are mutually exclusive combinations of CADGs which group patients with similarhealthcare needs. The grouping depends on ADGs and also on age and gender. The ACGclassifications were based on all patient diagnoses recorded over the one-year period prior toresource use as suggested by the system software (Johns Hopkins Bloomberg School of PublicHealth, 2008). Every patient was placed into one of 68 mutually exclusive ACGs.

4At least 35 of

these ACGs are for multimorbid patients in that they contain patients with at least 2 ADGs.

Resource Utilization Bands (RUBs). The ACG software groups ACGs with similar expectedexpenditure into 6 Resource Utilization Bands. 0 - No or only invalid diagnoses, 1 - Healthy Users, 2– Low, 3 – Moderate, 4 – High, 5 - Very High. Higher bands are expected to have higher costs andpatients in them are more likely to be multimorbid.

Table 2 summarises the three morbidity measures (vectors of QOF, Charlson, EDC morbiditymarkers) and the five measures of multimorbidity (counts of the QOF, Charlson, EDCs plus ACGs andRUBs) used in our analysis. The QOF, Charlson, and EDC morbidity categories were constructedusing all historic diagnoses on patients’ general practice records up to 31

stMarch 2007. The ACG and

RUB measures use diagnoses over the one-year period 1st

April 2006 to 31st

March 2007.

4Though the ACG System Version 8.2 identifies up to 82 default ACGs (or up to 93 if optional branching is turned on), the age

range of our sample meant that some of these categories were not populated.


2.1.4 Covariates

For each gender, age at 1st

April 2007 was categorised into ten-year age bands with the exception of90+ years as the upper category. Deprivation, measured by deciles of the Index of MultipleDeprivation (IMD) 2007 (Department for Communities and Local Government, 2008) which is basedon seven dimensions of deprivation, was attributed to the individuals by GPRD using patients’ LowerLayer Super Output Area (LSOA) of residence. There are 32,482 LSOAs in England with a meanpopulation of 1500.

2.2 Modelling

We estimated separate regression models of individual cost using the eight morbidity andmultimorbidity measures. The three numerical multimorbidity measures (QOF count, Charlson Indexscore, EDC count) and the ordered RUB multimorbidity measure were included as dummy variablecategories to estimate the most flexible relationships between multimorbidity and cost. In our dataset,the maximum QOF count was 10, the maximum Charlson score was 13, and the maximum EDCcount was 28. We used 6 categories for the QOF count (1,2,..,6 or more), 7 for the Charlson (1,2,…,7or more) and 18 for the EDC count (1,2,…,18 or more) as there were few patients with largernumerical scores. We used dummy variables for the 68 mutually exclusive ACG categories. For themodels with non-mutually exclusive QOF, Charlson, and EDC morbidity categories, we used dummyvariables for each of the categories.

The distribution of healthcare costs for individual patients usually has a long right hand tail and aspike at zero cost reflecting non-use by a non trivial proportion of the population. This has led tosome debate about the appropriate estimation method for models of individual cost with suggestionsincluding transformation of the cost variable, two part models, and Generalised Linear Models (GLMs)(Blough et al., 1999; Buntin and Zaslavsky, 2004; Manning, 2006; Manning and Mullahy, 2001;Manning et al., 2005; Mullahy, 1998). Our primary care cost data are right skewed (skewness 7.43,with the mean cost being 2.5 times the median), although the proportion of patients with no cost in2007/8 is smaller (12.3%) than in typical distributions of hospital costs.

We report results from estimated GLMs in which a link function of the conditional expected 2007/8cost of the i’th patient is linear in the explanatory variables:

0iadmp a ia d id m im p ipa d m pg Ec D D D D

(1)

and the error distributions were assumed to be Poisson (variance proportional to the mean), orGuassian (normal). The Dia are 15 age/gender group dummies, the Did are 9 deprivation deciledummies, the Dim are morbidity or multimorbidity category dummies, and the Dip are 173 practicedummies. All the explanatories are measured at the start of the expenditure year 2007/8 withmorbidity and multimorbidity variables based on patient morbidity records up to 31

stMarch 2007.

We used both log link functions (g(Ec) = ln Ec) and linear link functions (g(Ec) = Ec) The log formallows for the right skewness of the patient cost data and use of the GLM specification means that wedo not have to correct for retransformation bias (Manning 1998) or adjust the dependent variablebecause a proportion of patients have zero cost. With a linear link function and a normal errordistribution the GLM is equivalent to Ordinary Least Squares (OLS).

We estimated models with practice fixed effects Dip to allow for lack of information on practicecharacteristics (such as the number of GPs per patient) and for intangible “practice style” which mightaffect patient costs. Robust standard errors were used to allow for clustering of errors withinpractices. Models were estimated using STATA Version 11.2 (StataCorp, 2009).


We summarise model performance with four goodness of fit measures. The Bayesian InformationCriterion (BIC) (Schwarz, 1978) allows for differing numbers of explanatory variables in the modelsbeing compared and is calculated as

BIC = 2lnL*+ KlnN (2)

where L*

is the maximised likelihood from the model, N is the number of observations and K thenumber of parameters estimated.

The deviance-based2

DR for a model is

2 * 0 * 0 01 ln ln / ln ln ln ln / ln lnS S S

DR L L L L L L L L (3)

where L0

is the likelihood from a model with only a constant, L*

is the maximised likelihood for theestimated model, and L

Sis the likelihood from a saturated model with as many parameters as

observations.2

DR can be interpreted as the fraction of empirical uncertainty in total patient cost

which has been explained by the model (Cameron and Windmeijer, 1997). It is equal to the usual R2

in an OLS model (ie GLM with linear link and normal error distribution). Like the BIC its value dependson the assumed error distribution.

The Efron R2

(Efron, 1978) is the squared correlation coefficient between the estimated cost from amodel and actual cost and is readily obtained from an OLS regression of estimated on actual costs. Itis comparable across different underlying models for estimating cost and for OLS regression models it

is equal to the model R2

and to2

DR . We also computed the mean absolute error over all

observations (MAE), which is the average absolute difference in £’s between observed and estimatedcost. Like the Efron R

2it does not depend on the assumed error distribution and so can be used to

compare models with the same set of explanatories but different error distributions.


3. Results

3.1 Summary statistics

There were 86,100 individuals alive and registered at the beginning of the 2007/8 resource use year.Table 3 gives the frequency distribution by age and gender alongside the mean, standard deviationand median number of consultations, prescriptions, and tests per individual for 2007/8. Also shown isthe percentage of individuals who had at least one consultation, prescription, or test. At most ageswomen have higher rates of resource use but resource use increased more rapidly with age for menand was higher for men aged 80-89. The mean number of consultations and prescriptions is similarto those in other UK datasets for the same period (Hippisley-Cox and Vinogradova, 2009; NHSInformation Centre 2008b).

Table 3. Consultations, prescriptions and tests 2007/8: mean, (standard deviation), [median].

Agecategory N

Number ofconsultations

% at leastone

consultationNumber of

prescriptions

% at leastone

prescriptionNumber of

tests

% atleastonetest

Male 20-29 6021 1.7 (3.0) [1] 54 2.6 (10.0) [0] 44 0.7 (2.2) [0] 1930-39 7204 1.9 (3.5) [1] 54 4.1 (13.4) [0] 47 1.0 (2.9) [0] 22

40-49 8902 2.6 (4.4) [1] 60 6.6 (18.3) [1] 53 1.8 (4.1) [0] 30

50-59 7486 3.7 (5.4) [2] 69 13.4 (27.1) [2] 64 3.2 (5.7) [0] 44

60-69 6481 5.7 (6.9) [4] 83 26.5 (37.7) [15] 81 5.7 (7.9) [3] 64

70-79 4112 8.0 (7.8) [6] 92 42.9 (45.2) [33] 92 8.0 (8.9) [6] 77

80-89 1878 9.9 (9.7) [7] 93 54.4 (55.9) [44] 94 9.2 (10.2) [7] 79

90+ 253 7.3 (6.6) [6] 87 46.1 (52.0) [32] 88 6.0 (7.1) [4] 68

Female 20-29 5551 4.4 (4.8) [3] 84 5.8 (9.8) [3] 81 2.6 (4.6) [0] 4730-39 6930 4.6 (5.5) [3] 81 7.9 (20.4) [3] 76 2.9 (5.3) [1] 51

40-49 8447 4.5 (5.7) [3] 81 10.0 (23.6) [3] 74 3.1 (5.7) [1] 51

50-59 7525 5.0 (5.8) [3] 82 15.5 (28.2) [6] 78 4.2 (6.7) [1] 65

60-69 6563 6.3 (6.7) [5] 88 28.0 (40.5) [16] 87 5.9 (8.2) [3] 73

70-79 4954 8.3 (8.1) [6] 93 44.7 (50.8) [33] 93 7.8 (8.9) [6] 77

80-89 3057 9.1 (8.7) [7] 93 59.2 (68.0) [44] 95 7.9 (9.0) [6] 77

90+ 736 7.9 (7.8) [6] 90 64.6 (75.5) [47] 94 6.6 (7.8) [4] 72

Overall 86100 4.8 (6.3) [3] 77 18.5 (36.3) [4] 72 3.9 (6.8) [1] 52

Notes. Consultations are face to face or telephone consultations. 18.3% of patients had no consultation,prescription or test.

Table 4 gives the total cost per patient, stratified by age and gender, and categorised byconsultations, prescription drugs, and tests and investigations. Prescription drugs account for justover half (57%) of the total cost.

The mean total patient cost for men ranged from £90 at 20-29 years to £822 at 80-89 years. Forwomen the mean total patient cost ranged from £176 at 20-29 years to £709 at 80-89 years. Theshare of prescription drugs in total cost increased with age and was higher for men than for women.The share of consultations in total cost decreased with age, whereas the share of tests was fairlystable across age bands.

Deprivation has a weaker unconditional association with cost than age. The mean total patient costfor the most deprived decile is 1.34 times as large as mean total cost for the least deprived (£394 (CI:£379 to £408) compared to £289 (CI: £278 to £300)).


Table 4. Patient primary care costs (£) 2007/8: mean, (standard deviation), [median].Agecategory

Consultationcost

Prescription drugcost

Test andinvestigations cost

Total patient cost

Male 20-29 41 (69) [19] 43 (341) [0] 6 (32) [0] 90 (367) [24]30-39 48 (84) [21] 57 (248) [0] 9 (40) [0] 113 (300) [26]

40-49 61 (98) [26] 94 (477) [2] 13 (50) [0] 168 (524) [40]

50-59 86 (117) [49] 174 (443) [12] 22 (67) [0] 282 (530) [92]

60-69 130 (151) [93] 294 (503) [111] 34 (87) [8] 458 (622) [264]

70-79 184 (177) [139] 449 (602) [268] 49 (102) [14] 681 (722) [484]

80-89 237 (218) [178] 524 (643) [326] 60 (124) [17] 822 (783) [608]

90+ 213 (208) [160] 320 (381) [192] 34 (73) [10] 567 (510) [428]

Female 20-29 102 (110) [72] 57 (160) [17] 17 (50) [0] 176 (239) [106]30-39 106 (126) [72] 85 (330) [14] 20 (58) [0] 210 (407) [110]

40-49 105 (135) [68] 111 (324) [16] 24 (68) [0] 240 (429) [111]

50-59 119 (134) [80] 175 (418) [34] 40 (84) [11] 334 (515) [168]

60-69 148 (150) [109] 288 (533) [109] 47 (89) [16] 483 (636) [293]

70-79 195 (180) [150] 396 (541) [236] 48 (101) [14] 639 (662) [463]

80-89 234 (217) [179] 431 (665) [287] 45 (95) [13] 709 (772) [555]

90+ 245 (237) [189] 395 (434) [245] 32 (70) [10] 673 (579) [528]

Overall 114 (147) [69] 189 (458) [24] 27 (74) [0] 330 (563) [134]Notes. Consultation costs are the sum of the costs of face to face or telephone consultations plus the costs ofadministration for repeat prescriptions and other administration not requiring face to face or phone contact.12.3% of individuals had no cost.

Tables 5 and 6 have distributions of patients for each of the 8 morbidity and multimorbidity measures.The proportion of patients in the lowest multimorbidity categories is much smaller for the threemeasures (EDC count, ACG categories, RUB categories) derived by applying the ACG softwarewhich uses a much finer classification of morbidity than the 17 diseases in the QOF and Charlsonschemes. Although there are 68 mutually exclusive ACG categories, the eight largest categoriesaccounted for over 70% of the sample patients.

The distribution of the EDC count has a larger range than the QOF chronic disease count andCharlson Index score because of the greater number of relatively minor diseases that the EDC countincludes. According to the QOF chronic disease count 20% of patients were multimorbid (had a countof two or more) whereas 61% were multimorbid according to the EDC count. Women had slightlyhigher scores than men on the three count multimorbidity measures (QOF count, Charlson Indexscore, EDC count). There were significant positive Spearman rank correlations amongst these threecount measures (for the top censored counts used in the models) – QOF and Charlson: 0.63; QOFand EDC: 0.72; Charlson and EDC: 0.59.

There are differences across the QOF, Charlson and EDC morbidity categorisations in proportions ofpatients with some of the diseases. For example, 14.3% of patients have asthma in the EDC schemebut only 6.5% in the QOF scheme. The QOF payments for asthma patients relate mainly to themonitoring of patients and therefore patients require a recent inhaler prescription to be classified asasthmatic, whereas EDC requires only a diagnosis of asthma and therefore may capture all patientswho had childhood asthma. The QOF distinguishes between asthma and chronic obstructivepulmonary disease and so only records 2.1% of patients as having chronic obstructive pulmonarydisease, whereas the Charlson scheme records 16.6% as having chronic pulmonary disease becauseits definition includes asthmatics.


Table 5. Distribution of patients by QOF disease categories, QOF disease count, Charlson diseases, Charlson Index score, selected EDC categories, EDC count

QOF indicators % QOF count % Charlson diseases % Charlson score % Selected EDC categories % EDC count %

None 55.05 0 55.05 None 55.05 0 68.84 None 19.39 0 19.39

Asthma 6.5 1 24.81 AIDS 0.01 1 19.27 Low back pain 25.80 1 19.56

Atrial fibrillation 2.1 2 11.14 Cancer 3.83 2 6.62 Dermatitis and eczema 19.57 2 16.36

Cancer 1.6 3 5.20 Cerebrovascular disease 2.31 3 2.98 Hypertension 18.36 3 12.54

CHD 5.6 4 2.39 Chronic pulmonarydisease

16.63 4 1.30 Anxiety, neuroses 16.50 4 9.20

CKD 3.8 5 0.92 Congestive heartdisease

1.21 5 0.56 Depression 16.21 5 6.75

COPD 2.1 6+ 0.49 Dementia 0.41 6 0.23 Asthma 14.26 6 4.90

Dementia 0.5 Diabetes 4.28 7+ 0.19 Cervical pain syndromes 13.20 7 3.51

Depression 14.9 Diabetes withcomplications

0.91 Arthritis 11.17 8 2.41

Diabetes 5 Hemiplegia/paraplegia 0.18 Irritable bowel syndrome 6.78 9 1.74

Epilepsy 0.9 Metastatic tumour 0.12 Gastroesophageal syndrome 6.70 10 1.25

Heart failure 1.1 Mild liver disease 0.17 Acute myocardial infarction 5.84 11 0.82

Hypertension 18.2 Moderate or severe liverdisease

0.04 Malignant neoplasms of theskin

2.53 12 0.57

Learning diff. 0.4 Myocardial infarction 1.70 Malignant neoplasms, breast 1.09 13 0.36

Mental health 0.9 Peptic ulcer 2.15 Emphysema, chronicbronchitis, COPD

2.38 14 0.27

Obesity 9.9 Peripheral vasculardisease

1.43 15 0.15

Stroke 2.4 Renal disease 4.68 16 0.09

Hyperthyroidism 3.9 Rheumatological disease 2.05 17 0.07

18+ 0.08

Notes. CHD: coronary heart disease; CKD: chronic kidney disease; COPD: chronic obstructive pulmonary disease; AIDS: acquired immune deficiency syndrome.


Table 6. Distribution of patients by selected ACG categories and RUBs

Selected ACG categories % RUB %

Non-Users 9.44 Non-user 9.44

No Diagnosis or Only Unclassified Diagnosis 20.05 Healthy user 36.36

Preventive/Administrative 6.47 Low morbidity 25.03

Acute Minor, Age 6+ 9.63 Moderate 27.26

Chronic Medical: Stable 2.49 High 1.70

2-3 Other ADG Combinations, Age 35+ 9.98 Very high 0.20

4-5 Other ADG Combinations, Age 45+, no Major ADGs 1.83

4-5 Other ADG Combinations, Age 45+, 1 Major ADGs 2.28

6-9 Other ADG Combinations, Age 35+, 0-1 Major ADGs 1.52

10+ Other ADG Combinations, Age 18+, 2 Major ADGs 0.05

6-9 Other ADG Combinations, Males, Age 18 to 34, 1 Major ADGs 0.01

Note. ACGs: Adjusted Clinical Groups; ADGs: Aggregated Diagnosis Groups; used to construct ACGs;

RUB: Resource Utilization Band constructed by aggregation of ACGs.

3.2 Comparison of GLM specifications

For the regression modelling we dropped 154 individuals with missing deprivation data to leave anestimation sample of 85,946.

Table 7. Goodness of fit of alternative specifications of model for total patient cost

Log, Poission Linear, Gaussian (OLS)

Model specification BIC 2

DR Efron

R2

MAE BIC Efron R2

= 2

DR = R

2

MAE

Age, gender (Model 1) 38446223 0.21 0.13 285 1320939 0.13 285

Age, gender, and deprivation (Model 2) 38041309 0.22 0.13 283 1320555 0.13 284

Age, gender, and practice (Model 3) 37725187 0.22 0.14 282 1322027 0.14 283

Age, gender, deprivation, and practice (Model 4) 37522639 0.23 0.14 281 1321883 0.14 282

(Model 4) + QOF disease indicators 29339460 0.40 0.25 244 1302791 0.31 234

(Model 4) + QOF chronic disease count 28523441 0.42 0.29 239 1305547 0.29 240

(Model 4) + Charlson indicators 32546370 0.33 0.22 259 1310235 0.25 255

(Model 4) + Charlson Index score 32274547 0.34 0.23 258 1311982 0.23 260

(Model 4) + EDC indicators 26255449 0.46 0.29 231 1295660 0.37 222

(Model 4) + EDC count 26196861 0.46 0.32 229 1302546 0.31 236

(Model 4) + ACG 28694630 0.41 0.27 242 1308944 0.27 248

(Model 4) + RUB 30367472 0.38 0.24 250 1312522 0.23 259

Notes. BIC: Bayesian Information Criterion. Smaller BIC indicates better fit and is comparable for different

models with same error distribution. 2

DR : deviance based R

2, which is comparable for different models with same

error distribution. Efron R2: correlation coefficient from OLS regression of estimated cost on actual cost. For

OLS models the Efron R2, the deviance based R

2, and the model R

2are equal. MAE: mean absolute error

Table 7 has goodness of fit statistics (BIC, deviance-based R-squared values2

DR , Efron R

2, MAE)

for GLM models with log link and Poisson errors and OLS models (linear link, normal distribution) withvarious sets of covariates. For the models without any morbidity or multimorbidity measures the loglink Poisson models and OLS models have very similar performance in terms of MAE and Efron R

2for

any given set of covariates. The log link Poisson GLM has lower MAE and higher Efron R2

than OLSfor 5 of the 8 models with morbidity or multimorbidity measures. OLS does better when morbidity ismeasured by dummies for the Charlson diseases, QOF diseases and the EDC indicators.


3.3 Comparison of morbidity and multimorbidity measures

Table 7 shows that the inclusion of any measure of morbidity or multimorbidity boosts theperformance of the regression models considerably. For example, with age and gender groups,

deprivation deciles, and practice effects the2

DR for the log link Poisson GLM specification is 0.23.

Adding the set of Charlson indicators, the worst performing of the eight morbidity and multimorbidity

measures, to the model increases the2

DR to 0.34. Similar increases in performance are seen with

the OLS model.

For any given estimation method, the rankings of the eight morbidity and multimorbidity measures by

the BIC,2

DR , Efron R

2and MAE criteria are very similar. In the log link Poisson GLM specification,

the EDC count has the best performance on all goodness of fit statistics (provided we use the third

decimal place for the2

DR ) closely followed by the set of 114 EDC indicators. These two EDC based

measures are noticeably better than the QOF count, 68 ACG indicators, 17 QOF disease categories,and the 6 RUBs, which in turn are markedly better than the Charlson Index score and 17 Charlsondisease categories.

Under OLS estimation, the EDC indicators have the best performance followed by the QOF indicatorsand then the EDC count. The three sets of morbidity category dummies (EDC, QOF, Charlson)performed better than the corresponding count multimorbidity measures.

Tables 8 and 9 give cost ratios for the morbidity and multimorbidity measures from the log link

Poisson models. (OLS models yield similar results.) The log link GLM model estimates ln Ec x or exp( )Ec x . Since all explanatory variables are categorical, the cost or risk ratio for a variable

xk is the ratio of expected cost when xk = 1 to the expected cost when xk = 0 and is

exp( *1 )k j jj kx

/ exp( *0 )k j jj k

x

= expk

.

For the three count multimorbidity measures (QOF, Charlson, and EDC) and the two mutuallyexclusive multimorbidity categorisations (ACGs, RUBs) the cost ratio is relative to a zero count or tonon-users. For the three sets of non-mutually exclusive morbidity dummies (QOF diseases, Charlsondiseases, EDCs) the cost ratio is the cost of the disease relative to not having that disease, rather tonot having any disease. For example, a patient with an EDC count of 4 has an expected cost which is6.13 times as large as a patient with a zero EDC count. A patient diagnosed as having cancer underthe QOF scheme has an expected cost 1.74 times as large as a patient without cancer.

Amongst the QOF morbidity categories epilepsy is the chronic disease with the biggest relative effect(2.32 compared with no epilepsy), though only 0.9% of the sample have the condition. The mostcommon QOF condition is hypertension (18.2% of the sample) with a cost ratio of 1.42. All the QOFdisease cost ratios are significantly greater than 1 but their range is limited (1.09 to 2.32). The 17Charlson diseases also have a similarly limited range of cost ratios (0.98 to 2.43).

5Of the 114 EDCs

cost ratios, 82 are significant. They also have a limited range: all but 8 are less than 2.0 and thelargest (Transplant status) has a cost ratio of 3.6 but only 0.05% of patients are in this category).

The baseline ACG category is non-users and all other categories have a cost ratio in excess of 1.The results for selected ACGs in Table 9 suggest that in general patients with more morbidities havehigher costs. The fact that cost ratios increase with RUB levels also suggests that multimorbidpatients are more costly as patients in higher RUBs are more likely to be multimorbid.

5Patients with AIDS do not increase general practice costs significantly because in the UK their care and the prescribing of

AIDS drugs is generally managed in secondary care.


Table 8. Cost ratios for QOF disease categories, QOF disease count, Charlson disease categories, Charlson Index score, EDC count

QOF disease categories*

QOF disease count**

Charlson disease categories*

Charlson Index score**

EDC count**

Disease Costratio

95% CI Count Costratio

95% CI Disease Costratio

95% CI Score Costratio

95% CI Count Costratio

95% CI

Asthma 1.80 (1.74,1.86) 0 1.00 - AIDS 0.98 (0.51,1.88) 0 1.00 - 0 1.00 -

Atrial fibrillation 1.11 (1.06,1.17) 1 2.41 (2.34,2.48) Cancer 1.54 (1.48,1.60) 1 2.05 (2.00,2.10) 1 2.11 (2.01,2.21)

Cancer 1.74 (1.65,1.84) 2 3.80 (3.67,3.93) Cerebrovascular dis. 1.32 (1.25,1.39) 2 2.59 (2.50,2.68) 2 3.34 (3.20,3.48)

CHD 1.45 (1.41,1.50) 3 4.86 (4.68,5.04) Chronic pulmonarydis.

1.64 (1.61,1.68) 3 3.16 (3.02,3.31) 3 4.64 (4.44,4.85)

CKD 1.15 (1.10,1.20) 4 5.81 (5.55,6.07) Congestive heart dis. 1.24 (1.17,1.32) 4 3.60 (3.42,3.80) 4 6.13 (5.85,6.42)

COPD 1.64 (1.56,1.72) 5 7.27 (6.85,7.73) Dementia 1.29 (1.16,1.44) 5 3.86 (3.58,4.16) 5 7.73 (7.37,8.11)

Dementia 1.27 (1.15,1.41) 6+ 7.56 (7.04,8.11) Diabetes 1.98 (1.92,2.05) 6 4.13 (3.70,4.62) 6 9.08 (8.65,9.53)

Depression 1.53 (1.49,1.57) Diabetes with comp. 2.43 (2.27,2.61) 7+ 4.54 (3.98,5.18) 7 11.01 (10.40,11.66)

Diabetes 1.73 (1.67,1.79) Hemiplegia/paraplegia 1.86 (1.46,2.36) 8 12.16 (11.49,12.86)

Epilepsy 2.32 (2.13,2.53) Metastatic tumour 1.36 (1.07,1.71) 9 13.74 (12.92,14.61)

Heart failure 1.09 (1.02,1.17) Mild liver disease 1.39 (1.16,1.67) 10 15.02 (14.13,15.97)

Hypertension 1.42 (1.38,1.46) Mod/sev liver disease 2.11 (1.32,3.38) 11 16.96 (15.81,18.20)

Learning 1.70 (1.41,2.05) Myocardial infarction 1.42 (1.35,1.50) 12 17.18 (15.89,18.58)

Mental health 2.06 (1.88,2.26) Peptic ulcer 1.31 (1.25,1.38) 13 19.52 (17.87,21.32)

Obesity 1.30 (1.26,1.34) Periph vasculardisease

1.26 (1.20,1.33) 14 19.05 (17.17,21.13)

Stroke 1.22 (1.16,1.27) Renal disease 1.34 (1.28,1.40) 15 21.04 (18.65,23.73)

Hyperthyroidism 1.21 (1.16,1.26) Rheumatological dis. 1.46 (1.39,1.53) 16 24.69 (21.43,28.46)

17 24.00 (19.98,28.83)

18+ 27.60 (24.08,31.63)

Notes.*Cost ratios are the estimated costs for a patient with the relevant disease divided by the estimated cost for a patient without that disease.

**Cost ratios are the

estimated costs for a patient with the relevant count divided by the estimated cost for a patient with no disease (zero count). Estimates from GLM log Poisson model withage/gender, deprivation and practice effects. EDC: Expanded Diagnosis Cluster. CHD: coronary heart disease; CKD: chronic kidney disease; COPD: chronic obstructivepulmonary disease; AIDS: acquired immune deficiency syndrome.


Table 9. Cost ratios for selected EDC categories and for selected ACG categories and Resource Utilization Bands

Selected EDC categories*

Selected ACG categories**

Resource Utilization Bands**

Category Costratio

95% CI Category Costratio

95% CI Band Costratio

95% CI

Low back pain 1.10 (1.08,1.13) Non-Users 1.00 - Non-user 1.00 -

Dermatitis and eczema 1.12 (1.09,1.15) No Diagnosis or Only Unclassified Diagnosis 2.58 (2.40,2.77) Healthyuser

3.28 (3.06,3.51)

Hypertension 1.38 (1.35,1.42) Preventive/Administrative 4.04 (3.72,4.39) Low morbid 5.60 (5.22,6.01)

Anxiety, neuroses 1.21 (1.18,1.24) Acute Minor, Age 6+ 4.25 (3.95,4.57) Moderate 9.54 (8.90,10.23)

Depression 1.29 (1.26,1.33) Chronic Medical: Stable 7.06 (6.49,7.68) High 13.44 (12.36,14.62)

Asthma 1.50 (1.46,1.54) 2-3 Other ADG Combinations, Age 35+ 8.93 (8.31,9.60) Very high 16.02 (13.81,18.58)

Cervical pain syndromes 1.10 (1.07,1.12) 4-5 Other ADG Combinations, Age 45+, no Major ADGs 10.90 (10.05,11.82)

Arthritis 1.13 (1.09,1.16) 4-5 Other ADG Combinations, Age 45+, 1 Major ADG 12.50 (11.56,13.51)

Irritable bowel syndrome 1.19 (1.14,1.23) 6-9 Other ADG Combinations, Age 35+, 0-1 Major ADG 16.32 (15.03,17.71)

Gastroesophageal reflux 1.27 (1.23,1.31) 10+ Other ADG Combinations, Age 18+, 2 Major ADGs 21.75 (16.90,27.99)

Acute myocardial infarction 1.26 (1.22,1.30) 6-9 Other ADG Combinations, Male, Age 18 to 34, 1 MajorADG

46.78 (20.99,104.22)

Malignant neoplasm of the skin 1.07 (1.02,1.12)

Malignant neoplasms, breast 1.56 (1.46,1.68)

Emphysema, chron bronchitis,COPD

1.30 (1.24,1.36)

Notes.*Cost ratios for EDCs are the estimated costs for a patient with the relevant disease divided by the estimated cost for a patient without that disease.

**Cost ratios for

mutually exclusive ACG categories and RUBs are the estimated costs for a patient in the ACG or RUB with the relevant count divided by the estimated cost for a patient withno use. Estimates from GLM log Poisson model with age, gender, deprivation and practice effects. EDC: Expanded Diagnosis Cluster; ADGs: Aggregated Diagnosis Groups.


0300

600

900

1200

1500

Pre

dic

ted

mea

na

nn

ualp

atie

ntco

st(£

)

0 1 2 3 4 5 6+

QOF chronic disease count

0300

600

900

1200

1500

Pre

dic

ted

mea

na

nn

ualp

atie

ntco

st(£

)

0 1 2 3 4 5 6 7+

Charlson Index score

0500

1000

1500

2000

2500

Pre

dic

ted

mea

na

nn

ualp

atie

ntco

st(£

)

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18+

Expanded Diagnostic Cluster (EDC) count

0500

1000

1500

2000

2500

Pre

dic

ted

mea

na

nn

ualp

atie

ntco

st(£

)

Non-user Healthy user Low morbidity Moderate High Very high

Resource Utilization Band

The cost ratios for the three count measures increase with the counts (except for EDC counts 14 and17, which are slightly smaller, respectively, than EDC counts 13 and 16) implying, like the ACG andRUB results, that patients with more diseases have higher costs. We re-estimated the QOF countand EDC count log Poisson models with the counts and their squares rather than with categories forthe counts. The cost ratio for the squared QOF count was 0.921 (CI: 0.918-0.925) and for thesquared EDC count was 0.985 (CI: 0.985-0.986). This means that the proportionate effect of anadditional disease on the cost ratio declines with the number of diseases.

6This is confirmed by

inspection of the cost ratios on the counts. For the EDC count, having 2 diseases increases costs by3.34/2.11 = 1.58 times compared with having one disease, having 3 diseases increases costs by 1.39times compared to having two diseases, and so on until having 6 diseases increases costs by 1.17times compared to having 5 diseases. Since the cost ratios for the counts increase less thanproportionately with the counts, the level of cost increases less than exponentially with the number ofdiseases.

7Figure 1 plots predicted costs against the EDC count, QOF count, Charlson score and

RUB level and suggests a roughly linear, rather than exponential, effect of increasing multimorbidityon the level of costs.

Note: Estimated costs from log link, Poisson model with age/gender bands, deprivation and practice effects.

Figure 1. Estimated mean annual patient cost (with 95% confidence interval) for multimorbidity countsand Resource Utilization Bands

3.4 Effects of age, gender, deprivation and practice

Table 7 suggests that including fixed practice effects in the models did not contribute greatly to modelperformance. Adding practice effects to models with only age and gender or only age, gender and

deprivation increases model2

DR by at most 0.01. Dropping all the practice fixed effects from the full

EDC count model reduces the2

DR by 0.02.

6With n equal to the count of diseases, the estimated model is

2

0 1 2ln(E )c n n so that the cost (risk) ratio on the n

2

term is exp(2), and dln (Ec)/dn = 2 < 0 if exp(2) < 1.7

Let the cost with n diseases be C(n). Then if the cost ratios increase linearly with the count C(n)/C(0) = exp(n) so that

C(n)/C(n1) = exp() and C(n) = C(0)exp(n).


Table 10 has the cost ratios for age/gender categories and deprivation deciles from the log linkPoisson model with EDC counts. (Results with other morbidity or multimorbidity measures aresimilar.) The effects of age and gender are qualitatively similar to those in Table 5 but with smallerdifferences between men and women at any given age and with much less of an effect of age on cost.This suggests that older patients are more expensive mainly because they are sicker.

Table 10. Conditional effects of age, gender and deprivation.

Age category Cost ratio 95% CI Deprivation decile Cost ratio 95% CI

Male 20-29 1.00 - (least deprived) 1 1.00 -

30-39 1.17 (1.05,1.32) 2 1.04 (0.99,1.09)

40-49 1.48 (1.32,1.67) 3 1.07 (1.02,1.12)

50-59 1.90 (1.71,2.12) 4 1.05 (1.00,1.10)

60-69 2.22 (2.00,2.48) 5 1.05 (1.00,1.11)

70-79 2.39 (2.14,2.67) 6 1.07 (1.01,1.13)

80-89 2.38 (2.12,2.67) 7 1.13 (1.07,1.19)

90+ 1.81 (1.56,2.10) 8 1.08 (1.02,1.14)

Female 20-29 1.54 (1.39,1.72) 9 1.14 (1.08,1.20)

30-39 1.57 (1.40,1.75) (most deprived) 10 1.22 (1.15,1.29)

40-49 1.60 (1.44,1.79)

50-59 1.85 (1.66,2.06)

60-69 2.11 (1.89,2.35)

70-79 2.17 (1.94,2.42)

80-89 2.06 (1.84,2.30)

90+ 2.01 (1.77,2.27)

Notes. Cost ratios for age bands are the estimated costs for a patient in the relevant age category divided by theestimated cost for a male patient in the lowest age band. Cost ratios for deprivation deciles are estimated costsfor a patient in the relevant decile divided by the estimated cost for a patient in the lowest decile. Estimates arefrom GLM log Poisson model with EDC count, age/gender, deprivation and practice effects.

Table 7 shows that including deprivation deciles in the regression models in addition to age and

gender improves their fit modestly: the2

DR increases by 0.01, which is about the same as the

improvement due to adding practice effects. Once morbidity measures are included in the model thecontribution of deprivation is very small: dropping the deprivation deciles from the full log link Poisson

model with EDC counts does not change the2

DR (to 2 decimal places) and increases the BIC slightly

by 45711. By contrast dropping all the age/gender bands reduces the2

DR slightly to 0.45 and

increases the BIC by 851817.

The gradient of cost with respect to deprivation is flatter than with respect to age and after inclusion ofmorbidity measures the cost ratio between the 10

thand 1

stdeciles is reduced to 1.22 (CI: 1.15, 1.29).

3.5 Sensitivity analyses

There were some patients with extremely high costs: the highest four costs ranged from £15,128 to£27,810 compared to the median cost of £134. The use of log link in the GLM models reduced thisdiscrepancy considerably (the log of the highest cost was 10.233 compared to the log of the medianof 4.895). Results were not sensitive to dropping the patients above the 99

thcost centile (£2471 or

7.812 in logs). The2

DR of the log Poisson model with EDC counts increased to 0.48 compared 0.46

for the model with a full set of observations. There was a slight reduction in the cost ratios for higherEDC counts.

We also estimated GLM models with a log link and gamma error distribution (error varianceproportional to the square of the mean). In all cases, the log link gamma models had higher MAE and


smaller Efron R2

than the log link Poisson models. The Efron R2

was much smaller for the modelswith EDC indicators (0.03 vs 0.29) and QOF indicators (0.11 vs 0.25).

Our results were not sensitive to the method for calculating prescribing costs which are the largestpart of the costs. Using national average amounts per prescription rather than recorded amounts tocalculate total primary care costs reduced model performance somewhat but estimated costs ratioswere similar and the model rankings were unchanged.

4. Conclusion

4.1 Limitations

The GPRD practices included in our sample were all of “research standard”. One implication is thatthe recording of patient diagnoses may be less thorough at other UK general practices, particularly fordiagnoses which are not included in the QOF. Measures of morbidity or multimorbidity constructedfor practices with lower quality data may therefore be underestimates. Basing a capitation scheme onresults from high data quality practices could lead to lower capitation payments for practices with lessgood quality data. However, recording of QOF morbidity variables is less likely to show this kind ofvariation because they are used as part of a national pay for performance scheme.

It is possible that part of the observed association between morbidity and cost is because patientswho have, for unobserved reasons, a propensity to consult more frequently are more likely to acquiremore diagnoses and to be given more prescriptions. This effect is likely to be most marked formeasures which include a large number of diagnostic categories and may partly explain why EDCshave the strongest association with expenditure.

We have estimated drug costs using data on prescriptions issued by the practice. This willoverestimate costs because not all prescriptions issued to patients result in the dispensing ofmedicines. Patients must pay for dispensed prescriptions but because of exemptions from paymenton grounds of age, income, and health, over 90% of NHS prescriptions are dispensed without chargeto the patient. Thus there is little financial incentive for patients not to have prescriptions dispensed.One study estimated that 5.2% of prescriptions written in a general practice were not dispensed(Beardon et al., 1993). It therefore seems unlikely that there is any large upward bias in our estimatesof drug costs.

We included cost data for patients who were alive and registered with our sample practices on 1st

April 2007 at the start of the financial year 2007/8. We made no adjustment for patients whoderegistered from the practice list during the resource use year. Thus if patients moved to anotherpractice during the year their total primary care costs including those incurred in other practices wouldbe under recorded in our data. If the aim is to estimate future costs which will be incurred by thepractice in which the patient is initially registered this does not present a problem. On average,setting capitation fees based on our results would ensure that practices would be paid for costsactually incurred by patients on their list at the start of the year. A retrospective adjustment could bemade at the end of the financial year for costs attributable to newly registered patients. We also makeno adjustment for patients who die. Such patients may have higher costs but to the extent that themorbidity measures are correlated with mortality in the coming year the estimated effects of differenttypes of morbidity will include some of the differentially higher costs of patients at higher risk of death.

We had limited information on patient socio-economic circumstances because the data suppliers wereconcerned to ensure that there was no risk of identifying patients. Thus we had only a single measureof deprivation (IMD) which was attributed to patients by their small area of residence. The IMD hasbeen widely used in other studies and is intended to reflect a number of dimensions of deprivation(including income, education, and housing). But nevertheless it would have been useful to haveindividual level data on these patient characteristics and also on the ethnicity of patients which is notincluded in the IMD measure.


4.2 Discussion

4.2.1 Practice effects

The practice dummy variables pick up the effects of characteristics of practices such as the GP topatient ratio, idiosyncratic practice treatment styles, and differences in the practice means of bothobserved and unobserved patient characteristics. The small impact of practice dummies implies thatthe practice characteristics and mean patient characteristics are either offsetting, or that there is littlecross practice variation in these variables, or that they have little effect. An alternative possibility isthat ‘practice style’ manifests itself in differences in case finding and thus the number and type ofdiagnoses that are recorded for patients. Hence models of cost which include measures of morbiditywill not find any difference across practices. But we found that practice effects are not important inexplaining costs even when no account is taken of morbidity, suggesting that case finding does notexplain the small impact of practice effects on patient costs.

4.2.2 Age and gender

Age is one of the factors affecting capitation payments to general practices and our results show thatage does have a marked effect on expenditure, with a roughly five fold variation in expenditurebetween the lowest age band (20-29) and those aged 80-89 if no allowance is made for morbidity ormultimorbidity. However, once morbidity or multimorbidity is allowed for the variation in cost with ageis much less marked: those aged 80-89 have costs which are only about twice as large as those aged20-29, rather than five times as large. Thus a major reason that the elderly have higher costs is thatthey have greater morbidity.

4.2.3 Deprivation

Deprivation is often suggested as having a strong relationship with health status and general practiceworkload and until 2004 the capitation payments received by general practices depended on thedeprivation level of the areas of residence of their patients. The results here suggest there is positiverelationship between deprivation and primary care expenditure. The positive deprivation gradient isflatter than the age gradient and is reduced by allowing for patient morbidity, but it is still statisticallyand economically significant. Previous survey based studies have reported mixed results about therelationship between deprivation and general practice workload. Bago D’Uva (2005), using the BritishHousehold Panel Survey, found that patients with higher income had more consultations. Morris et al(2005) used data from the Health Survey for England which has more health measures and found anegative but insignificant association of income and higher social class with consultations. Dixon et al(2007) reviewed the evidence and suggest that the NHS is pro-poor in terms of the distribution of GPconsultations once allowance is made for morbidity, age and gender. Our results, based on muchmore detailed health data than earlier studies support this broad conclusion. We cannot however ruleout the possibility that poorer health status in more deprived groups may be relatively under-recorded.

4.2.4 Estimation methods

With any of the morbidity or multimorbidity measures, both the log Poisson GLM and OLS estimationmethods had good explanatory power, especially for individual level cross-section regressions. OLSperformed a little worse with five of the eight morbidity and multimorbidity measures. This wasunexpected because in large samples of patients it has been found that OLS performs as well orbetter than GLM models in explaining hospital costs (Gravelle et al., 2011; Van de Ven and Ellis,2000) and our sample of primary care patients had a smaller proportion of patients with noexpenditure than in these studies. However, the reduction in explanatory power was not large andOLS estimation with the 114 EDC categories had the lowest MAE and highest Efron R

2over all sets

of explanatories and estimation methods. The estimation method was less important than the choiceof morbidity or multimorbidity measure.

4.2.5 Morbidity and multimorbidity measures

Having finer categories of morbidity unsurprisingly improved the ability to predict patient costs. Thusmeasures using the 114 EDC categories were better than those using the 17 category QOF scheme,whether the categories were used to measure morbidity as a set of dummy variables or used toconstruct counts of disease categories to measure multimorbidity. The QOF based measures ofmorbidity and multimorbidity performed considerably better than the Charlson based measures whichhad the same number of morbidity categories. The poor performance of the widely used CharlsonIndex score and of the Charlson disease dummy variables may be a reflection of the fact that the


Charlson scheme was originally intended to predict mortality rather than the cost of general practiceactivities. The two QOF based measures had about the same predictive power as the 68 mutuallyexclusive ACG categories derived using purpose built case-mix software. This may be because the17 QOF categories were selected for a primary care pay for performance scheme targeted at care forchronic patients who are the main business of general practices. The ACG categories included non-chronic diagnoses which were grouped in part by their anticipated effect on all patient costs, includinghospital costs.

Measuring morbidity as a set of disease dummy variables and measuring multimorbidity as a count ofdiseases are both simple ways of using the rich information available in electronic patient healthrecords. Their simplicity requires possibly incorrect assumptions about the relationship betweenmorbidity and cost. With linear models of cost, using morbidity categories allows the effect of costs tovary across diseases but assumes that there are no interactions between diseases in their effect oncost.

8If there are non-additive effects of multimorbidity then these will be partially picked up in the

estimated coefficients on the morbidity category dummies since they estimate the average cost ofpatients with only that disease plus some of the effect on costs of patients with that and otherdiseases.

9On the other hand, using a count of diseases with dummies for each count value allows

flexibly for non-additivity but assumes that all diseases are essentially identical in their effect on cost.For example, the coefficient on the first count is the average cost of all patients with a single disease,the coefficient on the second count is the average cost of all patients with two diseases, and so on.Thus the choice between using disease dummies to measure morbidity or a count of the samediseases to measure multimorbidity reflects a trade-off between two possible sources of bias inestimated cost function coefficients. The fact that the Charlson disease, QOF disease, and EDCindicator dummies have better MAE and Efron R

2in the linear OLS models than the corresponding

count multimorbidity measures suggests that it is more important to allow for different diseases tohave different cost impacts than to allow for possible non-additivity. However, the difference betweenthe multimorbidity counts and the corresponding morbidity categories in terms of ability to explainpatient costs is not great. This may be due to the fact that costs increased roughly linearly with thenumber of diseases, and there was relatively little variation in costs of particular diseases.

4.2.6 Practice budget setting

We included measures based on the 17 chronic QOF conditions because such information is readilyavailable and reliably coded in all UK general practices. The measures performed nearly as well asmeasures requiring considerably more information. As the QOF includes major chronic diseases it islikely that similar measures could be readily constructed using routine data in other countries.Moreover, it would currently be feasible to use the results from the linear OLS model with a simple setof 17 QOF disease dummies to allocate budgets to practices to cover their primary care costs. Withan additive linear model, practice budgets can be set without information on the QOF diseases of allindividual patients. For a practice level allocation the coefficients on the QOF diseases can beapplied to the number of a practice’s patients with each of the 17 QOF diseases. Data on thenumbers with each disease in each practice is already produced by the QOF process. Similarlyapplying the deprivation and age/gender coefficients to numbers on practice lists in each deprivationand age/gender category yields an estimate of total practice primary care needs.

10Such a procedure

seems likely to improve on the current methods of allocating funds to practices to cover their costs ofproviding primary care.

8The discussion is in terms of linear cost models but with suitable reinterpretation it applies to log models. For example

additivity in the log model means that the proportional effects on cost (rather than absolute effects) of having a disease do notdepend on whether the patient has other diseases.9

Suppose there are two diseases with prevalence i (unconditional probability), with 1 + 2 < 1 so that there are some non-

morbid patients, and that the patient cost is c = 0 + 1D1 + 2D2 + 12D1D2 + where Di = 1 if and only if the individual has

disease i. If the diseases are statistically independent and we estimate a model c = a0 + b1D1 + b2D2 + e, then plim ib = i +

j12, where j is the probability of disease j. Even in this simple case we also get an inconsistent estimate of the cost of

multimorbid patient with both diseases by adding the separate coefficients: plim ( 1b + 2b ) = 1 + 2 + (1+2) 12 ≠ 1 + 2 + 12.10

Assumes that variations in practice effects reflect differences in practice behaviour rather than patient need and should nottherefore determine the budget, the expenditure which is associated with needs variables in practice p is

0ˆ ˆ ˆ ˆ

p a pa d pd m pma d mn n n n

where np is the total list, and npa, npd, npm are the number of patients in the practice in age/gender band a, deprivation decile d,and with QOF disease m.


4.2.7 Simpler is better?

Our study is the first to compare across a range of estimation methods and measures of morbidityand multimorbidity with respect to their ability to explain primary health care resource use in the UK.The best performing multimorbidity measures were simple counts of the number of chronic conditionspatients suffered from or simple sets of disease dummies. In future work it would be useful to testwhether other reasonably straightforward specifications would improve the ability to predict patientcosts. One interesting, and simple, possibility is to add indicators for particular combinations ofdiseases to models.


References

Bago d'Uva, T. 2005. Latent class models for use of primary care: evidence from a British panel.Health Economics. 14, 873-892.

Bago d'Uva, T., Jones, A.M., van Doorslaer, E., 2009. Measurement of horizontal inequity in healthcare utilisation using European panel data, Journal of Health Economics, 28, 280-289.

Beardon P, McGilchrist M, McKendrick A, McDevitt D, MacDonald T. 1993. Primary non-compliancewith prescribed medication in primary care. British Medical Journal. 307: 846-848.

British Medical Association and Royal Pharmaceutical Society. 2010. http://bnf.org. Accessed 12 May2011.

Blough, D.K., Madden, C.W., Hornbrook, M.C. 1999. Modeling risk using generalized linear models.Journal of Health Economics, 18: 153-171.

Buntin, M.B., Zaslavsky, A.M. 2004. Too much ado about two-part models and transformation?Comparing methods of modelling Medicare expenditures. Journal of Health Economics, 23: 525-542.

Cameron, A.C., Windmeijer, F.A.G. 1997. An R-squared measure of goodness of fit for somecommon nonlinear regression models. Journal of Econometrics, 77, (2) 329-342.

Charlson, M.E., Pompei, P., Ales, K.L., & Mackenzie, C.R. 1987. A new method of classifyingprognostic co-morbidity in longitudinal-studies - development and validation. Journal of ChronicDiseases, 40, (5) 373-383.

Curtis, L. 2008. Unit costs of health and social care 2007 PSSRU, University of Kent.Department for Communities and Local Government. The English Indices of Deprivation 2007. 28-3-2008.

Department of Health. 2009. “NHS Reference Costs 2007/08”. [online]http://www.dh.gov.uk/en/Publicationsandstatistics/Publications/PublicationsLibrary/index.htm.Accessed 10 May 2011.

Department of Health. 2010. Equity and excellence: Liberating the NHS. Cm 7881.

Department of Health. 2011. Resource Allocation: Weighted Capitation Formula. 7th

Edition.http://www.dh.gov.uk/prod_consum_dh/groups/dh_digitalassets/documents/digitalasset/dh_124947.pdf. Accessed 5 September 2011.

Deyo, R.A., Cherkin, D.C., Ciol, M.A. 1992. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. Journal of Clinical Epidemiology, 45, (6) 613-619.

Dixon, A., Le Grand, J., Henderson, J., Murray, R., Poteliakhoff, E. 2007. Is the British NationalHealth Service equitable? The evidence on socioeconomic differences in utilization. Journal of HealthServices Research and Policy. 12, 2, 104-109.

Efron, B. 1978. Regression and ANOVA with zero-one data: measures of residual variation. Journalof the American Statistical Association. 73, 113-121.

Formula Review Group. 2007. Review of the General Medical Services global sum formula. BritishMedical Association and NHS Employers. http://www.nhsemployers.org/SiteCollectionDocuments/frg_report_final_cd_090207.pdf Accessed 5 September 2011.

Gravelle, H., Dusheiko, M., Martin, S., Rice, N., Smith, P., Dixon, J. 2011. Modelling individual patienthospital expenditure for general practice budgets. CHE Research Paper 72. Centre for HealthEconomics, University of York.


Gravelle, H., Morris, S., Sutton, M. 2006. Economic studies of equity in the consumption of health care.In Jones, A.M. (ed.), Elgar Companion to Health Economics, Edward Elgar,193-204.

Halling, A., Fridh, G., & Ovhed, I. 2006. Validating the Johns Hopkins ACG case-mix system of theelderly in Swedish primary health care. Bmc Public Health, 6., available from: ISI:000239803600001

Hippisley-Cox J, Pringle M. 2007. Comorbidity of diseases in the new General Medical ServicesContract for General Practitioners: analysis of QRESEARCH data.http://www.qresearch.org/Public_Documents/DataValidation/Co-morbidity%20of%20diseases%20in%20the%20new%20GMS%20contract%20for%20GPs.pdf

Hippisley-Cox, J. & Vinogradova, Y. 2009. Trends in consultation rates in general practice 1995 to2008: Analysis of the QResearch database. Final Report to the NHS Information Centre andDepartment of Health. National Health Service (NHS) Information Centre and QResearch.

Huntley, A., Johnson, R., Purdy, S., Valderas, J., Salisbury, C. 2012. Measures of multimorbidity andmorbidity burden for use in primary care settings: a systematic review and guide. Annals of FamilyMedicine. (In press)

Johns Hopkins Bloomberg School of Public Health. 2008. The Johns Hopkins ACG® Case-mixSystem Version 8.2. Baltimore,

Khan, N.F., Perera, R., Harper, S., & Rose, P.W. 2010. Adaptation and validation of the CharlsonIndex for Read/OXMIS coded databases. BMC Family Practice, 11, available from:ISI:000274624100001

Lawrenson, R., Williams, T., & Farmer, R. 1999. Clinical information for research; the use of generalpractice databases. Journal of Public Health Medicine, 21, (3) 299-304.

Manning, W.G. 1998. The logged dependent variable, heteroskedasticity, and the retransformationproblem. Journal of Health Economics, 17: 283-295.

Manning, W.G. 2006. Dealing with skewed data on costs and expenditures. In Jones, A.M. (ed.),Elgar Companion to Health Economics, Edward Elgar, 439-446.

Manning, W.G., Mullahy, J. 2001. Estimating log models: to transform or not to transform? Journal ofHealth Economics, 20: 461-494.

Manning, W.G., Basu, A., Mullahy, J. 2005. Generalised modelling approaches to risk adjustment ofskewed outcomes data. Journal of Health Economics. 24. 465-488.

Morris, S., Sutton, M., Gravelle, H. 2005. Inequity and inequality in the use of health care in England:an empirical investigation. Social Science and Medicine, 60, 1251-1266.

Mullahy, J. 1998. Much ado about two: reconsidering retransformation and the two-part model inhealth econometrics. Journal of Health Economics, 17: 247-281.

National Health Service (NHS) Information Centre. 2008a. Prescription cost analysis 2007.

National Health Service (NHS) Information Centre. 2008b. Prescriptions dispensed in the community,statistics for 1997 to 2007: England.

Omar, R.Z., O'Sullivan, C., Petersen, I., Islam, A., & Majeed, A. 2008. A model based on age, sex,and morbidity to explain variation in UK general practice prescribing: cohort study. British MedicalJournal, 337, doi:10.1136/bmj.a238

Perkins, A.J., Kroenke, K., Unutzer, J., Katon, W., Williams, J.W., Hope, C., & Callahan, C.M. 2004.Common comorbidity scales were similar in their ability to predict health care costs and mortality.Journal of Clinical Epidemiology, 57, (10) 1040-1048.


Reid, R.J., Roos, N.P., MacWilliam, L., Frohlich, N., & Black, C. 2002. Assessing population healthcare need using a claims-based ACG morbidity measure: A validation analysis in the province ofManitoba. Health Services Research, 37, (5) 1345-1364.

Salisbury, C., Johnson, L., Purdy, S., Valderas, J.M., & Montgomery, A.A. 2011. Epidemiology andimpact of multimorbidity in primary care: a retrospective cohort study. British Journal of GeneralPractice, 61, (582) 18-24.

Schokkaert, E., Dhaene, G., and Van de Voorde, C. 1998. Risk adjustment and the trade-off betweenefficiency and risk selection: an application of the theory of fair compensation. Health Economics7:465-480.

Schwarz, G. 1978. Estimating dimension of a model. Annals of Statistics, 6, (2) 461-464.

Starfield, B., Weiner, J., Mumford, L., & Steinwachs, D. 1991. Ambulatory care groups - Acategorization of diagnoses for research and management. Health Services Research, 26, (1) 53-74.

StataCorp. 2009. Stata Statistical Software: Release 11. College Station, TX: StataCorp LP.

Sullivan, C.O., Omar, R.Z., Ambler, G., & Majeed, A. 2005. Case-mix and variation in specialistreferrals in general practice. British Journal of General Practice, 55, (516) 529-533.

Sutton, M., 2002. Vertical and horizontal aspects of socio-economic inequity in general practitionercontacts in Scotland. Health Economics. 11, 537–549.

Technical Steering Committee 2010, GP Earnings and expenses 2007/08. NHS Information Centre.

van de Ven, W. P. and R. P. Ellis (2000): Risk adjustment in competitive health plan markets,Elsevier, vol. 1 of Handbook of Health Economics, 755–845.

Weiner, J.P., Starfield, B.H., Steinwachs, D.M., & Mumford, L.M. 1991. Development and applicationof a population-oriented measure of ambulatory care case-mix. Medical Care, 29, (5) 452-472.

Winkelman, R., Mehmud, S. 2007. A comparative analysis of claims based tools for health riskassessment. Society of Actuaries. www.soa.org/files/pdf/risk-assessmentc.pdf. Accessed 18October 2011.